Report-wide selection terms

Report-wide transforms operate on one evaluated value per occurrence, after the caller's explicit grouping and prediction model selection. Raw duplicate model rows do not add observations. One evaluation partition must contain only one named sample; evaluate reports separately or supply separate policy source contexts. A policy's pre-filter runs before its scoring population is formed.

from topiary import Column, ordinal_rank
support = Column("support_reads").log2()
lens_support = support / support.population_max()
expression_rank = Column("allele_expression").dense_rank(0, 1)
order = ordinal_rank(Column("tier_order"), expression_rank, Column("gene"))

The same expressions parse from strings and survive policy evidence replay. population_max() broadcasts the maximum finite value, leaving an empty or entirely missing population unknown. It does not add pseudocounts or decide what zero-denominator normalization should mean.

dense_rank(ascending, missing_bottom) starts at 1 and gives equal values equal ranks. Both arguments are 0/1; defaults are ascending with missing ranks kept missing. Explicit bottom ranking places all missing/nonfinite values after finite values. ordinal_rank(key, ...) ranks ascending lexicographic tuples, including text Column keys; missing keys sort last and exact ties remain tied. Contradictory per-occurrence column values raise.

expression.fillna(value) explicitly fills missing numeric values. For an arbitrary Column it also defines the absent-column fallback. This is useful when CCF is genuinely optional, not permission to fabricate required evidence. Ambiguous models, conflicting values and nonfinite domain errors stay visible.

These are generic ranking primitives, not fitted scientific models. Consumers must still pin published definitions, evidence units, source parameters and normalization populations. See openvax/topiary#462 and openvax/vaxrank#497.